root / tcg / optimize.c @ fa01a208
History | View | Annotate | Download (22.7 kB)
| 1 |
/*
|
|---|---|
| 2 |
* Optimizations for Tiny Code Generator for QEMU
|
| 3 |
*
|
| 4 |
* Copyright (c) 2010 Samsung Electronics.
|
| 5 |
* Contributed by Kirill Batuzov <batuzovk@ispras.ru>
|
| 6 |
*
|
| 7 |
* Permission is hereby granted, free of charge, to any person obtaining a copy
|
| 8 |
* of this software and associated documentation files (the "Software"), to deal
|
| 9 |
* in the Software without restriction, including without limitation the rights
|
| 10 |
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
| 11 |
* copies of the Software, and to permit persons to whom the Software is
|
| 12 |
* furnished to do so, subject to the following conditions:
|
| 13 |
*
|
| 14 |
* The above copyright notice and this permission notice shall be included in
|
| 15 |
* all copies or substantial portions of the Software.
|
| 16 |
*
|
| 17 |
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
| 18 |
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
| 19 |
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
|
| 20 |
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
| 21 |
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
| 22 |
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
|
| 23 |
* THE SOFTWARE.
|
| 24 |
*/
|
| 25 |
|
| 26 |
#include "config.h" |
| 27 |
|
| 28 |
#include <stdlib.h> |
| 29 |
#include <stdio.h> |
| 30 |
|
| 31 |
#include "qemu-common.h" |
| 32 |
#include "tcg-op.h" |
| 33 |
|
| 34 |
#define CASE_OP_32_64(x) \
|
| 35 |
glue(glue(case INDEX_op_, x), _i32): \
|
| 36 |
glue(glue(case INDEX_op_, x), _i64)
|
| 37 |
|
| 38 |
typedef enum { |
| 39 |
TCG_TEMP_UNDEF = 0,
|
| 40 |
TCG_TEMP_CONST, |
| 41 |
TCG_TEMP_COPY, |
| 42 |
TCG_TEMP_HAS_COPY, |
| 43 |
TCG_TEMP_ANY |
| 44 |
} tcg_temp_state; |
| 45 |
|
| 46 |
struct tcg_temp_info {
|
| 47 |
tcg_temp_state state; |
| 48 |
uint16_t prev_copy; |
| 49 |
uint16_t next_copy; |
| 50 |
tcg_target_ulong val; |
| 51 |
}; |
| 52 |
|
| 53 |
static struct tcg_temp_info temps[TCG_MAX_TEMPS]; |
| 54 |
|
| 55 |
/* Reset TEMP's state to TCG_TEMP_ANY. If TEMP was a representative of some
|
| 56 |
class of equivalent temp's, a new representative should be chosen in this
|
| 57 |
class. */
|
| 58 |
static void reset_temp(TCGArg temp, int nb_temps, int nb_globals) |
| 59 |
{
|
| 60 |
int i;
|
| 61 |
TCGArg new_base = (TCGArg)-1;
|
| 62 |
if (temps[temp].state == TCG_TEMP_HAS_COPY) {
|
| 63 |
for (i = temps[temp].next_copy; i != temp; i = temps[i].next_copy) {
|
| 64 |
if (i >= nb_globals) {
|
| 65 |
temps[i].state = TCG_TEMP_HAS_COPY; |
| 66 |
new_base = i; |
| 67 |
break;
|
| 68 |
} |
| 69 |
} |
| 70 |
for (i = temps[temp].next_copy; i != temp; i = temps[i].next_copy) {
|
| 71 |
if (new_base == (TCGArg)-1) { |
| 72 |
temps[i].state = TCG_TEMP_ANY; |
| 73 |
} else {
|
| 74 |
temps[i].val = new_base; |
| 75 |
} |
| 76 |
} |
| 77 |
temps[temps[temp].next_copy].prev_copy = temps[temp].prev_copy; |
| 78 |
temps[temps[temp].prev_copy].next_copy = temps[temp].next_copy; |
| 79 |
} else if (temps[temp].state == TCG_TEMP_COPY) { |
| 80 |
temps[temps[temp].next_copy].prev_copy = temps[temp].prev_copy; |
| 81 |
temps[temps[temp].prev_copy].next_copy = temps[temp].next_copy; |
| 82 |
new_base = temps[temp].val; |
| 83 |
} |
| 84 |
temps[temp].state = TCG_TEMP_ANY; |
| 85 |
if (new_base != (TCGArg)-1 && temps[new_base].next_copy == new_base) { |
| 86 |
temps[new_base].state = TCG_TEMP_ANY; |
| 87 |
} |
| 88 |
} |
| 89 |
|
| 90 |
static int op_bits(TCGOpcode op) |
| 91 |
{
|
| 92 |
const TCGOpDef *def = &tcg_op_defs[op];
|
| 93 |
return def->flags & TCG_OPF_64BIT ? 64 : 32; |
| 94 |
} |
| 95 |
|
| 96 |
static TCGOpcode op_to_movi(TCGOpcode op)
|
| 97 |
{
|
| 98 |
switch (op_bits(op)) {
|
| 99 |
case 32: |
| 100 |
return INDEX_op_movi_i32;
|
| 101 |
case 64: |
| 102 |
return INDEX_op_movi_i64;
|
| 103 |
default:
|
| 104 |
fprintf(stderr, "op_to_movi: unexpected return value of "
|
| 105 |
"function op_bits.\n");
|
| 106 |
tcg_abort(); |
| 107 |
} |
| 108 |
} |
| 109 |
|
| 110 |
static void tcg_opt_gen_mov(TCGArg *gen_args, TCGArg dst, TCGArg src, |
| 111 |
int nb_temps, int nb_globals) |
| 112 |
{
|
| 113 |
reset_temp(dst, nb_temps, nb_globals); |
| 114 |
assert(temps[src].state != TCG_TEMP_COPY); |
| 115 |
if (src >= nb_globals) {
|
| 116 |
assert(temps[src].state != TCG_TEMP_CONST); |
| 117 |
if (temps[src].state != TCG_TEMP_HAS_COPY) {
|
| 118 |
temps[src].state = TCG_TEMP_HAS_COPY; |
| 119 |
temps[src].next_copy = src; |
| 120 |
temps[src].prev_copy = src; |
| 121 |
} |
| 122 |
temps[dst].state = TCG_TEMP_COPY; |
| 123 |
temps[dst].val = src; |
| 124 |
temps[dst].next_copy = temps[src].next_copy; |
| 125 |
temps[dst].prev_copy = src; |
| 126 |
temps[temps[dst].next_copy].prev_copy = dst; |
| 127 |
temps[src].next_copy = dst; |
| 128 |
} |
| 129 |
gen_args[0] = dst;
|
| 130 |
gen_args[1] = src;
|
| 131 |
} |
| 132 |
|
| 133 |
static void tcg_opt_gen_movi(TCGArg *gen_args, TCGArg dst, TCGArg val, |
| 134 |
int nb_temps, int nb_globals) |
| 135 |
{
|
| 136 |
reset_temp(dst, nb_temps, nb_globals); |
| 137 |
temps[dst].state = TCG_TEMP_CONST; |
| 138 |
temps[dst].val = val; |
| 139 |
gen_args[0] = dst;
|
| 140 |
gen_args[1] = val;
|
| 141 |
} |
| 142 |
|
| 143 |
static TCGOpcode op_to_mov(TCGOpcode op)
|
| 144 |
{
|
| 145 |
switch (op_bits(op)) {
|
| 146 |
case 32: |
| 147 |
return INDEX_op_mov_i32;
|
| 148 |
case 64: |
| 149 |
return INDEX_op_mov_i64;
|
| 150 |
default:
|
| 151 |
fprintf(stderr, "op_to_mov: unexpected return value of "
|
| 152 |
"function op_bits.\n");
|
| 153 |
tcg_abort(); |
| 154 |
} |
| 155 |
} |
| 156 |
|
| 157 |
static TCGArg do_constant_folding_2(TCGOpcode op, TCGArg x, TCGArg y)
|
| 158 |
{
|
| 159 |
switch (op) {
|
| 160 |
CASE_OP_32_64(add): |
| 161 |
return x + y;
|
| 162 |
|
| 163 |
CASE_OP_32_64(sub): |
| 164 |
return x - y;
|
| 165 |
|
| 166 |
CASE_OP_32_64(mul): |
| 167 |
return x * y;
|
| 168 |
|
| 169 |
CASE_OP_32_64(and): |
| 170 |
return x & y;
|
| 171 |
|
| 172 |
CASE_OP_32_64(or): |
| 173 |
return x | y;
|
| 174 |
|
| 175 |
CASE_OP_32_64(xor): |
| 176 |
return x ^ y;
|
| 177 |
|
| 178 |
case INDEX_op_shl_i32:
|
| 179 |
return (uint32_t)x << (uint32_t)y;
|
| 180 |
|
| 181 |
case INDEX_op_shl_i64:
|
| 182 |
return (uint64_t)x << (uint64_t)y;
|
| 183 |
|
| 184 |
case INDEX_op_shr_i32:
|
| 185 |
return (uint32_t)x >> (uint32_t)y;
|
| 186 |
|
| 187 |
case INDEX_op_shr_i64:
|
| 188 |
return (uint64_t)x >> (uint64_t)y;
|
| 189 |
|
| 190 |
case INDEX_op_sar_i32:
|
| 191 |
return (int32_t)x >> (int32_t)y;
|
| 192 |
|
| 193 |
case INDEX_op_sar_i64:
|
| 194 |
return (int64_t)x >> (int64_t)y;
|
| 195 |
|
| 196 |
case INDEX_op_rotr_i32:
|
| 197 |
x = ((uint32_t)x << (32 - y)) | ((uint32_t)x >> y);
|
| 198 |
return x;
|
| 199 |
|
| 200 |
case INDEX_op_rotr_i64:
|
| 201 |
x = ((uint64_t)x << (64 - y)) | ((uint64_t)x >> y);
|
| 202 |
return x;
|
| 203 |
|
| 204 |
case INDEX_op_rotl_i32:
|
| 205 |
x = ((uint32_t)x << y) | ((uint32_t)x >> (32 - y));
|
| 206 |
return x;
|
| 207 |
|
| 208 |
case INDEX_op_rotl_i64:
|
| 209 |
x = ((uint64_t)x << y) | ((uint64_t)x >> (64 - y));
|
| 210 |
return x;
|
| 211 |
|
| 212 |
CASE_OP_32_64(not): |
| 213 |
return ~x;
|
| 214 |
|
| 215 |
CASE_OP_32_64(neg): |
| 216 |
return -x;
|
| 217 |
|
| 218 |
CASE_OP_32_64(andc): |
| 219 |
return x & ~y;
|
| 220 |
|
| 221 |
CASE_OP_32_64(orc): |
| 222 |
return x | ~y;
|
| 223 |
|
| 224 |
CASE_OP_32_64(eqv): |
| 225 |
return ~(x ^ y);
|
| 226 |
|
| 227 |
CASE_OP_32_64(nand): |
| 228 |
return ~(x & y);
|
| 229 |
|
| 230 |
CASE_OP_32_64(nor): |
| 231 |
return ~(x | y);
|
| 232 |
|
| 233 |
CASE_OP_32_64(ext8s): |
| 234 |
return (int8_t)x;
|
| 235 |
|
| 236 |
CASE_OP_32_64(ext16s): |
| 237 |
return (int16_t)x;
|
| 238 |
|
| 239 |
CASE_OP_32_64(ext8u): |
| 240 |
return (uint8_t)x;
|
| 241 |
|
| 242 |
CASE_OP_32_64(ext16u): |
| 243 |
return (uint16_t)x;
|
| 244 |
|
| 245 |
case INDEX_op_ext32s_i64:
|
| 246 |
return (int32_t)x;
|
| 247 |
|
| 248 |
case INDEX_op_ext32u_i64:
|
| 249 |
return (uint32_t)x;
|
| 250 |
|
| 251 |
default:
|
| 252 |
fprintf(stderr, |
| 253 |
"Unrecognized operation %d in do_constant_folding.\n", op);
|
| 254 |
tcg_abort(); |
| 255 |
} |
| 256 |
} |
| 257 |
|
| 258 |
static TCGArg do_constant_folding(TCGOpcode op, TCGArg x, TCGArg y)
|
| 259 |
{
|
| 260 |
TCGArg res = do_constant_folding_2(op, x, y); |
| 261 |
if (op_bits(op) == 32) { |
| 262 |
res &= 0xffffffff;
|
| 263 |
} |
| 264 |
return res;
|
| 265 |
} |
| 266 |
|
| 267 |
static TCGArg do_constant_folding_cond(TCGOpcode op, TCGArg x,
|
| 268 |
TCGArg y, TCGCond c) |
| 269 |
{
|
| 270 |
switch (op_bits(op)) {
|
| 271 |
case 32: |
| 272 |
switch (c) {
|
| 273 |
case TCG_COND_EQ:
|
| 274 |
return (uint32_t)x == (uint32_t)y;
|
| 275 |
case TCG_COND_NE:
|
| 276 |
return (uint32_t)x != (uint32_t)y;
|
| 277 |
case TCG_COND_LT:
|
| 278 |
return (int32_t)x < (int32_t)y;
|
| 279 |
case TCG_COND_GE:
|
| 280 |
return (int32_t)x >= (int32_t)y;
|
| 281 |
case TCG_COND_LE:
|
| 282 |
return (int32_t)x <= (int32_t)y;
|
| 283 |
case TCG_COND_GT:
|
| 284 |
return (int32_t)x > (int32_t)y;
|
| 285 |
case TCG_COND_LTU:
|
| 286 |
return (uint32_t)x < (uint32_t)y;
|
| 287 |
case TCG_COND_GEU:
|
| 288 |
return (uint32_t)x >= (uint32_t)y;
|
| 289 |
case TCG_COND_LEU:
|
| 290 |
return (uint32_t)x <= (uint32_t)y;
|
| 291 |
case TCG_COND_GTU:
|
| 292 |
return (uint32_t)x > (uint32_t)y;
|
| 293 |
} |
| 294 |
break;
|
| 295 |
case 64: |
| 296 |
switch (c) {
|
| 297 |
case TCG_COND_EQ:
|
| 298 |
return (uint64_t)x == (uint64_t)y;
|
| 299 |
case TCG_COND_NE:
|
| 300 |
return (uint64_t)x != (uint64_t)y;
|
| 301 |
case TCG_COND_LT:
|
| 302 |
return (int64_t)x < (int64_t)y;
|
| 303 |
case TCG_COND_GE:
|
| 304 |
return (int64_t)x >= (int64_t)y;
|
| 305 |
case TCG_COND_LE:
|
| 306 |
return (int64_t)x <= (int64_t)y;
|
| 307 |
case TCG_COND_GT:
|
| 308 |
return (int64_t)x > (int64_t)y;
|
| 309 |
case TCG_COND_LTU:
|
| 310 |
return (uint64_t)x < (uint64_t)y;
|
| 311 |
case TCG_COND_GEU:
|
| 312 |
return (uint64_t)x >= (uint64_t)y;
|
| 313 |
case TCG_COND_LEU:
|
| 314 |
return (uint64_t)x <= (uint64_t)y;
|
| 315 |
case TCG_COND_GTU:
|
| 316 |
return (uint64_t)x > (uint64_t)y;
|
| 317 |
} |
| 318 |
break;
|
| 319 |
} |
| 320 |
|
| 321 |
fprintf(stderr, |
| 322 |
"Unrecognized bitness %d or condition %d in "
|
| 323 |
"do_constant_folding_cond.\n", op_bits(op), c);
|
| 324 |
tcg_abort(); |
| 325 |
} |
| 326 |
|
| 327 |
|
| 328 |
/* Propagate constants and copies, fold constant expressions. */
|
| 329 |
static TCGArg *tcg_constant_folding(TCGContext *s, uint16_t *tcg_opc_ptr,
|
| 330 |
TCGArg *args, TCGOpDef *tcg_op_defs) |
| 331 |
{
|
| 332 |
int i, nb_ops, op_index, nb_temps, nb_globals, nb_call_args;
|
| 333 |
TCGOpcode op; |
| 334 |
const TCGOpDef *def;
|
| 335 |
TCGArg *gen_args; |
| 336 |
TCGArg tmp; |
| 337 |
/* Array VALS has an element for each temp.
|
| 338 |
If this temp holds a constant then its value is kept in VALS' element.
|
| 339 |
If this temp is a copy of other ones then this equivalence class'
|
| 340 |
representative is kept in VALS' element.
|
| 341 |
If this temp is neither copy nor constant then corresponding VALS'
|
| 342 |
element is unused. */
|
| 343 |
|
| 344 |
nb_temps = s->nb_temps; |
| 345 |
nb_globals = s->nb_globals; |
| 346 |
memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info)); |
| 347 |
|
| 348 |
nb_ops = tcg_opc_ptr - gen_opc_buf; |
| 349 |
gen_args = args; |
| 350 |
for (op_index = 0; op_index < nb_ops; op_index++) { |
| 351 |
op = gen_opc_buf[op_index]; |
| 352 |
def = &tcg_op_defs[op]; |
| 353 |
/* Do copy propagation */
|
| 354 |
if (!(def->flags & (TCG_OPF_CALL_CLOBBER | TCG_OPF_SIDE_EFFECTS))) {
|
| 355 |
assert(op != INDEX_op_call); |
| 356 |
for (i = def->nb_oargs; i < def->nb_oargs + def->nb_iargs; i++) {
|
| 357 |
if (temps[args[i]].state == TCG_TEMP_COPY) {
|
| 358 |
args[i] = temps[args[i]].val; |
| 359 |
} |
| 360 |
} |
| 361 |
} |
| 362 |
|
| 363 |
/* For commutative operations make constant second argument */
|
| 364 |
switch (op) {
|
| 365 |
CASE_OP_32_64(add): |
| 366 |
CASE_OP_32_64(mul): |
| 367 |
CASE_OP_32_64(and): |
| 368 |
CASE_OP_32_64(or): |
| 369 |
CASE_OP_32_64(xor): |
| 370 |
CASE_OP_32_64(eqv): |
| 371 |
CASE_OP_32_64(nand): |
| 372 |
CASE_OP_32_64(nor): |
| 373 |
if (temps[args[1]].state == TCG_TEMP_CONST) { |
| 374 |
tmp = args[1];
|
| 375 |
args[1] = args[2]; |
| 376 |
args[2] = tmp;
|
| 377 |
} |
| 378 |
break;
|
| 379 |
CASE_OP_32_64(brcond): |
| 380 |
if (temps[args[0]].state == TCG_TEMP_CONST |
| 381 |
&& temps[args[1]].state != TCG_TEMP_CONST) {
|
| 382 |
tmp = args[0];
|
| 383 |
args[0] = args[1]; |
| 384 |
args[1] = tmp;
|
| 385 |
args[2] = tcg_swap_cond(args[2]); |
| 386 |
} |
| 387 |
break;
|
| 388 |
CASE_OP_32_64(setcond): |
| 389 |
if (temps[args[1]].state == TCG_TEMP_CONST |
| 390 |
&& temps[args[2]].state != TCG_TEMP_CONST) {
|
| 391 |
tmp = args[1];
|
| 392 |
args[1] = args[2]; |
| 393 |
args[2] = tmp;
|
| 394 |
args[3] = tcg_swap_cond(args[3]); |
| 395 |
} |
| 396 |
break;
|
| 397 |
CASE_OP_32_64(movcond): |
| 398 |
if (temps[args[1]].state == TCG_TEMP_CONST |
| 399 |
&& temps[args[2]].state != TCG_TEMP_CONST) {
|
| 400 |
tmp = args[1];
|
| 401 |
args[1] = args[2]; |
| 402 |
args[2] = tmp;
|
| 403 |
args[5] = tcg_swap_cond(args[5]); |
| 404 |
} |
| 405 |
default:
|
| 406 |
break;
|
| 407 |
} |
| 408 |
|
| 409 |
/* Simplify expressions for "shift/rot r, 0, a => movi r, 0" */
|
| 410 |
switch (op) {
|
| 411 |
CASE_OP_32_64(shl): |
| 412 |
CASE_OP_32_64(shr): |
| 413 |
CASE_OP_32_64(sar): |
| 414 |
CASE_OP_32_64(rotl): |
| 415 |
CASE_OP_32_64(rotr): |
| 416 |
if (temps[args[1]].state == TCG_TEMP_CONST |
| 417 |
&& temps[args[1]].val == 0) { |
| 418 |
gen_opc_buf[op_index] = op_to_movi(op); |
| 419 |
tcg_opt_gen_movi(gen_args, args[0], 0, nb_temps, nb_globals); |
| 420 |
args += 3;
|
| 421 |
gen_args += 2;
|
| 422 |
continue;
|
| 423 |
} |
| 424 |
break;
|
| 425 |
default:
|
| 426 |
break;
|
| 427 |
} |
| 428 |
|
| 429 |
/* Simplify expression for "op r, a, 0 => mov r, a" cases */
|
| 430 |
switch (op) {
|
| 431 |
CASE_OP_32_64(add): |
| 432 |
CASE_OP_32_64(sub): |
| 433 |
CASE_OP_32_64(shl): |
| 434 |
CASE_OP_32_64(shr): |
| 435 |
CASE_OP_32_64(sar): |
| 436 |
CASE_OP_32_64(rotl): |
| 437 |
CASE_OP_32_64(rotr): |
| 438 |
CASE_OP_32_64(or): |
| 439 |
CASE_OP_32_64(xor): |
| 440 |
if (temps[args[1]].state == TCG_TEMP_CONST) { |
| 441 |
/* Proceed with possible constant folding. */
|
| 442 |
break;
|
| 443 |
} |
| 444 |
if (temps[args[2]].state == TCG_TEMP_CONST |
| 445 |
&& temps[args[2]].val == 0) { |
| 446 |
if ((temps[args[0]].state == TCG_TEMP_COPY |
| 447 |
&& temps[args[0]].val == args[1]) |
| 448 |
|| args[0] == args[1]) { |
| 449 |
gen_opc_buf[op_index] = INDEX_op_nop; |
| 450 |
} else {
|
| 451 |
gen_opc_buf[op_index] = op_to_mov(op); |
| 452 |
tcg_opt_gen_mov(gen_args, args[0], args[1], |
| 453 |
nb_temps, nb_globals); |
| 454 |
gen_args += 2;
|
| 455 |
} |
| 456 |
args += 3;
|
| 457 |
continue;
|
| 458 |
} |
| 459 |
break;
|
| 460 |
default:
|
| 461 |
break;
|
| 462 |
} |
| 463 |
|
| 464 |
/* Simplify expression for "op r, a, 0 => movi r, 0" cases */
|
| 465 |
switch (op) {
|
| 466 |
CASE_OP_32_64(and): |
| 467 |
CASE_OP_32_64(mul): |
| 468 |
if ((temps[args[2]].state == TCG_TEMP_CONST |
| 469 |
&& temps[args[2]].val == 0)) { |
| 470 |
gen_opc_buf[op_index] = op_to_movi(op); |
| 471 |
tcg_opt_gen_movi(gen_args, args[0], 0, nb_temps, nb_globals); |
| 472 |
args += 3;
|
| 473 |
gen_args += 2;
|
| 474 |
continue;
|
| 475 |
} |
| 476 |
break;
|
| 477 |
default:
|
| 478 |
break;
|
| 479 |
} |
| 480 |
|
| 481 |
/* Simplify expression for "op r, a, a => mov r, a" cases */
|
| 482 |
switch (op) {
|
| 483 |
CASE_OP_32_64(or): |
| 484 |
CASE_OP_32_64(and): |
| 485 |
if (args[1] == args[2]) { |
| 486 |
if (args[1] == args[0]) { |
| 487 |
gen_opc_buf[op_index] = INDEX_op_nop; |
| 488 |
} else {
|
| 489 |
gen_opc_buf[op_index] = op_to_mov(op); |
| 490 |
tcg_opt_gen_mov(gen_args, args[0], args[1], nb_temps, |
| 491 |
nb_globals); |
| 492 |
gen_args += 2;
|
| 493 |
} |
| 494 |
args += 3;
|
| 495 |
continue;
|
| 496 |
} |
| 497 |
break;
|
| 498 |
default:
|
| 499 |
break;
|
| 500 |
} |
| 501 |
|
| 502 |
/* Propagate constants through copy operations and do constant
|
| 503 |
folding. Constants will be substituted to arguments by register
|
| 504 |
allocator where needed and possible. Also detect copies. */
|
| 505 |
switch (op) {
|
| 506 |
CASE_OP_32_64(mov): |
| 507 |
if ((temps[args[1]].state == TCG_TEMP_COPY |
| 508 |
&& temps[args[1]].val == args[0]) |
| 509 |
|| args[0] == args[1]) { |
| 510 |
args += 2;
|
| 511 |
gen_opc_buf[op_index] = INDEX_op_nop; |
| 512 |
break;
|
| 513 |
} |
| 514 |
if (temps[args[1]].state != TCG_TEMP_CONST) { |
| 515 |
tcg_opt_gen_mov(gen_args, args[0], args[1], |
| 516 |
nb_temps, nb_globals); |
| 517 |
gen_args += 2;
|
| 518 |
args += 2;
|
| 519 |
break;
|
| 520 |
} |
| 521 |
/* Source argument is constant. Rewrite the operation and
|
| 522 |
let movi case handle it. */
|
| 523 |
op = op_to_movi(op); |
| 524 |
gen_opc_buf[op_index] = op; |
| 525 |
args[1] = temps[args[1]].val; |
| 526 |
/* fallthrough */
|
| 527 |
CASE_OP_32_64(movi): |
| 528 |
tcg_opt_gen_movi(gen_args, args[0], args[1], nb_temps, nb_globals); |
| 529 |
gen_args += 2;
|
| 530 |
args += 2;
|
| 531 |
break;
|
| 532 |
CASE_OP_32_64(not): |
| 533 |
CASE_OP_32_64(neg): |
| 534 |
CASE_OP_32_64(ext8s): |
| 535 |
CASE_OP_32_64(ext8u): |
| 536 |
CASE_OP_32_64(ext16s): |
| 537 |
CASE_OP_32_64(ext16u): |
| 538 |
case INDEX_op_ext32s_i64:
|
| 539 |
case INDEX_op_ext32u_i64:
|
| 540 |
if (temps[args[1]].state == TCG_TEMP_CONST) { |
| 541 |
gen_opc_buf[op_index] = op_to_movi(op); |
| 542 |
tmp = do_constant_folding(op, temps[args[1]].val, 0); |
| 543 |
tcg_opt_gen_movi(gen_args, args[0], tmp, nb_temps, nb_globals);
|
| 544 |
} else {
|
| 545 |
reset_temp(args[0], nb_temps, nb_globals);
|
| 546 |
gen_args[0] = args[0]; |
| 547 |
gen_args[1] = args[1]; |
| 548 |
} |
| 549 |
gen_args += 2;
|
| 550 |
args += 2;
|
| 551 |
break;
|
| 552 |
CASE_OP_32_64(add): |
| 553 |
CASE_OP_32_64(sub): |
| 554 |
CASE_OP_32_64(mul): |
| 555 |
CASE_OP_32_64(or): |
| 556 |
CASE_OP_32_64(and): |
| 557 |
CASE_OP_32_64(xor): |
| 558 |
CASE_OP_32_64(shl): |
| 559 |
CASE_OP_32_64(shr): |
| 560 |
CASE_OP_32_64(sar): |
| 561 |
CASE_OP_32_64(rotl): |
| 562 |
CASE_OP_32_64(rotr): |
| 563 |
CASE_OP_32_64(andc): |
| 564 |
CASE_OP_32_64(orc): |
| 565 |
CASE_OP_32_64(eqv): |
| 566 |
CASE_OP_32_64(nand): |
| 567 |
CASE_OP_32_64(nor): |
| 568 |
if (temps[args[1]].state == TCG_TEMP_CONST |
| 569 |
&& temps[args[2]].state == TCG_TEMP_CONST) {
|
| 570 |
gen_opc_buf[op_index] = op_to_movi(op); |
| 571 |
tmp = do_constant_folding(op, temps[args[1]].val,
|
| 572 |
temps[args[2]].val);
|
| 573 |
tcg_opt_gen_movi(gen_args, args[0], tmp, nb_temps, nb_globals);
|
| 574 |
gen_args += 2;
|
| 575 |
} else {
|
| 576 |
reset_temp(args[0], nb_temps, nb_globals);
|
| 577 |
gen_args[0] = args[0]; |
| 578 |
gen_args[1] = args[1]; |
| 579 |
gen_args[2] = args[2]; |
| 580 |
gen_args += 3;
|
| 581 |
} |
| 582 |
args += 3;
|
| 583 |
break;
|
| 584 |
CASE_OP_32_64(setcond): |
| 585 |
if (temps[args[1]].state == TCG_TEMP_CONST |
| 586 |
&& temps[args[2]].state == TCG_TEMP_CONST) {
|
| 587 |
gen_opc_buf[op_index] = op_to_movi(op); |
| 588 |
tmp = do_constant_folding_cond(op, temps[args[1]].val,
|
| 589 |
temps[args[2]].val, args[3]); |
| 590 |
tcg_opt_gen_movi(gen_args, args[0], tmp, nb_temps, nb_globals);
|
| 591 |
gen_args += 2;
|
| 592 |
} else {
|
| 593 |
reset_temp(args[0], nb_temps, nb_globals);
|
| 594 |
gen_args[0] = args[0]; |
| 595 |
gen_args[1] = args[1]; |
| 596 |
gen_args[2] = args[2]; |
| 597 |
gen_args[3] = args[3]; |
| 598 |
gen_args += 4;
|
| 599 |
} |
| 600 |
args += 4;
|
| 601 |
break;
|
| 602 |
CASE_OP_32_64(brcond): |
| 603 |
if (temps[args[0]].state == TCG_TEMP_CONST |
| 604 |
&& temps[args[1]].state == TCG_TEMP_CONST) {
|
| 605 |
if (do_constant_folding_cond(op, temps[args[0]].val, |
| 606 |
temps[args[1]].val, args[2])) { |
| 607 |
memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info)); |
| 608 |
gen_opc_buf[op_index] = INDEX_op_br; |
| 609 |
gen_args[0] = args[3]; |
| 610 |
gen_args += 1;
|
| 611 |
} else {
|
| 612 |
gen_opc_buf[op_index] = INDEX_op_nop; |
| 613 |
} |
| 614 |
} else {
|
| 615 |
memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info)); |
| 616 |
reset_temp(args[0], nb_temps, nb_globals);
|
| 617 |
gen_args[0] = args[0]; |
| 618 |
gen_args[1] = args[1]; |
| 619 |
gen_args[2] = args[2]; |
| 620 |
gen_args[3] = args[3]; |
| 621 |
gen_args += 4;
|
| 622 |
} |
| 623 |
args += 4;
|
| 624 |
break;
|
| 625 |
CASE_OP_32_64(movcond): |
| 626 |
if (temps[args[1]].state == TCG_TEMP_CONST |
| 627 |
&& temps[args[2]].state == TCG_TEMP_CONST) {
|
| 628 |
tmp = do_constant_folding_cond(op, temps[args[1]].val,
|
| 629 |
temps[args[2]].val, args[5]); |
| 630 |
if (args[0] == args[4-tmp] |
| 631 |
|| (temps[args[4-tmp]].state == TCG_TEMP_COPY
|
| 632 |
&& temps[args[4-tmp]].val == args[0])) { |
| 633 |
gen_opc_buf[op_index] = INDEX_op_nop; |
| 634 |
} else if (temps[args[4-tmp]].state == TCG_TEMP_CONST) { |
| 635 |
gen_opc_buf[op_index] = op_to_movi(op); |
| 636 |
tcg_opt_gen_movi(gen_args, args[0], temps[args[4-tmp]].val, |
| 637 |
nb_temps, nb_globals); |
| 638 |
gen_args += 2;
|
| 639 |
} else {
|
| 640 |
gen_opc_buf[op_index] = op_to_mov(op); |
| 641 |
tcg_opt_gen_mov(gen_args, args[0], args[4-tmp], |
| 642 |
nb_temps, nb_globals); |
| 643 |
gen_args += 2;
|
| 644 |
} |
| 645 |
} else {
|
| 646 |
reset_temp(args[0], nb_temps, nb_globals);
|
| 647 |
gen_args[0] = args[0]; |
| 648 |
gen_args[1] = args[1]; |
| 649 |
gen_args[2] = args[2]; |
| 650 |
gen_args[3] = args[3]; |
| 651 |
gen_args[4] = args[4]; |
| 652 |
gen_args[5] = args[5]; |
| 653 |
gen_args += 6;
|
| 654 |
} |
| 655 |
args += 6;
|
| 656 |
break;
|
| 657 |
case INDEX_op_call:
|
| 658 |
nb_call_args = (args[0] >> 16) + (args[0] & 0xffff); |
| 659 |
if (!(args[nb_call_args + 1] & (TCG_CALL_CONST | TCG_CALL_PURE))) { |
| 660 |
for (i = 0; i < nb_globals; i++) { |
| 661 |
reset_temp(i, nb_temps, nb_globals); |
| 662 |
} |
| 663 |
} |
| 664 |
for (i = 0; i < (args[0] >> 16); i++) { |
| 665 |
reset_temp(args[i + 1], nb_temps, nb_globals);
|
| 666 |
} |
| 667 |
i = nb_call_args + 3;
|
| 668 |
while (i) {
|
| 669 |
*gen_args = *args; |
| 670 |
args++; |
| 671 |
gen_args++; |
| 672 |
i--; |
| 673 |
} |
| 674 |
break;
|
| 675 |
default:
|
| 676 |
/* Default case: we do know nothing about operation so no
|
| 677 |
propagation is done. We trash everything if the operation
|
| 678 |
is the end of a basic block, otherwise we only trash the
|
| 679 |
output args. */
|
| 680 |
if (def->flags & TCG_OPF_BB_END) {
|
| 681 |
memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info)); |
| 682 |
} else {
|
| 683 |
for (i = 0; i < def->nb_oargs; i++) { |
| 684 |
reset_temp(args[i], nb_temps, nb_globals); |
| 685 |
} |
| 686 |
} |
| 687 |
for (i = 0; i < def->nb_args; i++) { |
| 688 |
gen_args[i] = args[i]; |
| 689 |
} |
| 690 |
args += def->nb_args; |
| 691 |
gen_args += def->nb_args; |
| 692 |
break;
|
| 693 |
} |
| 694 |
} |
| 695 |
|
| 696 |
return gen_args;
|
| 697 |
} |
| 698 |
|
| 699 |
TCGArg *tcg_optimize(TCGContext *s, uint16_t *tcg_opc_ptr, |
| 700 |
TCGArg *args, TCGOpDef *tcg_op_defs) |
| 701 |
{
|
| 702 |
TCGArg *res; |
| 703 |
res = tcg_constant_folding(s, tcg_opc_ptr, args, tcg_op_defs); |
| 704 |
return res;
|
| 705 |
} |